Scott Alexander

It has dimensions for real money, easy to use and easy to create market.

And all of the vertices on the cube have been marked with different models or different platforms that seem to correspond with that.

So one vertex here is totally useless.

All three of the aspects are bad.

FTX is good for money.

Metaculous is good for easy to use.

And foretold is good for easy to create market.

That's the three lines that lead from that vertex.

Also labeled we have polymarket, which is good for money and easy to use.

We have manifold, which is good for easy to use and easy to create market.

But the vertex diagonally opposite, totally useless, which has all three, is labeled with a question mark.

When I ask Manifold why they won't add one, they say that Polymarket and Cauchy already dominate the space, and they have other, more interesting plans to be announced soon.

When I ask Polymarket why they won't do two, the answer is a combination of regulatory issues, fear that people would write bad resolution criteria and it would reflect badly on them, and there always being something more important to do.

I haven't asked Cauchy, but their answer would definitely be regulatory.

I still think this is a billion dollar bill waiting to be picked up.

The second is to conclude that prediction markets' role in God's plan was only to provide the foundation for AI superforecasters, the training data, the benchmarking arena, and the pot of money that rewards innovation.

Once AI superforecasters are developed, then, for all that the rest of us care, the markets themselves can wither away into the sports gambling casinos they so desperately want to become.

The Forecasting Research Institute's linear extrapolation shows AIs are on track to match top humans, quote, by late 2026.

Here's a graph showing improving performance as different models are released.

So the dates are on the x-axis from July 2023 to July 2025.